Shanghai-based Fysics AI launched the Fysiverse on June 25, a model that hardcodes real-world physics laws into its architecture. Founder Zhang Lihua, a former senior manager at Nvidia, is proposing a frontal alternative to the data-driven paths of Sora and V-JEPA. The bet: make AI reliable in situations video models have never seen.
Key Takeaways
- Fysics AI (Shanghai) ships the Fysiverse, a world model that embeds physics laws in its code
- The approach breaks from OpenAI’s Sora and Meta’s V-JEPA, both data-driven
- Target markets: content creation, robotic training, autonomous driving
A New Way to Build a World Model
The Fysiverse was unveiled on June 25 through Fysics AI’s WeChat account. The pitch is short: a model that adheres to real physics laws. The difference is a choice of architecture, not of scale. Where Western frontier models learn physics from video, Fysics AI encodes it into the model itself.
A world model produces internally consistent simulations: a ball that falls bounces, a spilled liquid flows, a heavy object deforms what it touches. For a classic LLM or video model, that consistency depends entirely on the size and quality of the training set.
Fysics AI attacks that point directly. The company says its rivals leak three families of errors: physical illusions (a scene that looks fine but breaks gravity or inertia), causal reasoning failures, and breakdowns in non-standard scenarios, the very ones missing from training video.
Founder Zhang Lihua comes from Nvidia, where he was a senior manager. That profile matters: the team has the GPU culture and the CUDA fluency needed to run a model under hard physics constraints. This is not a fine-tuning layer on top of an open-source base, it is a different object.
The move fits a broader Chinese trajectory. Over the past year DeepSeek has shaken the frontier race with model-side efficiency, and DeepSeek raised 50 billion dollars with Tencent to fund the next scale. Fysics AI is playing a different card: architecture, rather than raw data volume.
Sora, V-JEPA, Fysiverse: Three Paths, One Goal
Comparing the three approaches sharpens the bet. Sora, at OpenAI, learns from massive video datasets. The model infers regularities of the world by statistical imitation: the more clips it sees, the more it “knows” how a ball rolls on a table.
V-JEPA at Meta takes a different route. It uses self-supervised learning to reconstruct what is missing from a scene, without ever receiving a physics textbook. Implicit understanding is meant to emerge from the volume of observed data.
Fysiverse chooses the opposite. Physics laws are set as upstream constraints, encoded into the architecture. The network learns to reason inside a space where gravity, inertia and energy conservation are not something to discover but something to obey. The promise: fewer illusions, fewer failures in edge cases.
This is not the first time China has tried to short-circuit the American roadmap through architecture. The pattern was already visible with open-weight models, where MiniMax M3 shipped a competitive open-weight model against closed offerings from OpenAI and Anthropic. Fysics AI is aiming for the same shift, on another segment.
The proof question remains open. Fysics AI is publishing a claim and a target use case. Public benchmarks are not yet available. A world model that says it outperforms on non-standard scenarios still has to show it on evaluations a third party can replicate.
Also on Horizon:
- Zuckerberg AI Agents Slower Than Expected at Meta
- Sakana AI: We Tested the Japanese AI Everyone Talks About
- SpaceX AI Phone Prototype Leaks, Musk Denies It All
Three Target Markets: Robotics, Driving, Content
Fysics AI lists three markets. The first is content creation, with fewer physics glitches in generated video and animation. The second is robot training. A robot trained in a simulator that obeys real physics laws transfers its skills into the real world with fewer surprises.
The third is autonomous driving. A self-driving car has to anticipate rare situations (icy braking, a pedestrian stepping behind an obstacle, a car reacting on a deformed road surface). These scenarios are underrepresented in training videos, precisely where Fysiverse claims an edge.
Short term, the impact plays out on the Chinese market. Domestic players (automakers, video platforms, industrial robotics companies) structurally prefer local building blocks. Fysics AI does not need to convince San Francisco to land its first big contracts.
Medium term, the question shifts to export. A Chinese model with a strong robotics angle falls into the categories Washington monitors closely. The customer-by-customer reviews of frontier models show the US administration is now drawing very fine lines around the movement of sensitive AI bricks.
A complementary signal comes from enterprise. For six months, US players have been testing Chinese AI to cut their compute bill. Coinbase halved its AI bill by migrating to Chinese models. If Fysiverse holds up technically, that logic can extend into the world-model segment.
The real unknown is the timing of any opening. Fysics AI has not said whether Fysiverse will be offered as an API, an enterprise license, or an open-weight release. Each of the three options would radically change adoption speed and the company’s standing in the landscape.
Follow the story on Horizon.
